this post was submitted on 16 Jul 2023
85 points (91.3% liked)
Asklemmy
43947 readers
598 users here now
A loosely moderated place to ask open-ended questions
Search asklemmy ๐
If your post meets the following criteria, it's welcome here!
- Open-ended question
- Not offensive: at this point, we do not have the bandwidth to moderate overtly political discussions. Assume best intent and be excellent to each other.
- Not regarding using or support for Lemmy: context, see the list of support communities and tools for finding communities below
- Not ad nauseam inducing: please make sure it is a question that would be new to most members
- An actual topic of discussion
Looking for support?
Looking for a community?
- Lemmyverse: community search
- sub.rehab: maps old subreddits to fediverse options, marks official as such
- [email protected]: a community for finding communities
~Icon~ ~by~ ~@Double_[email protected]~
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
What makes you think that? This is wrong. Sure you can try and train a neuronal network to remember something exactly. But this would waste gigabytes of memory and lots of computing for some photo that you could just store on the smallest thumbdrive as a jpg and clone it with the digital precision, computers are made for. You don't need a neural net for that. And once you start feeding it the third or fourth photo, the first one will deteriorate and it will become difficult to reproduce each of them exactly. I'm not an expert on machine learning, but i think the fact that floating point arithmetic has a certain, finite precision and we're talking about statistics and hundreds of thousands to millions of pixels per photo makes it even more difficult to store things exactly.
Actually the way machine learning models work is: It has a look at lots of photos and each time adapts its weights a tiny bit. Nothing gets copied 1:1. A small amount if information is transferred from the item into the weights. And that is the way you want it to work to be useful. It should not memorise each of van gogh's paintings 1:1 because this wouldn't allow you to create a new fake van gogh. You want it to understand how van gogh's style looks. You want it to learn concepts and store more abstract knowledge, that it can then apply to new tasks. I hope i explained this well enough. If machine learning worked the way you described, it would be nothing more than expensive storage. It could reproduce things 1:1 but you obviously can't tell your thumbdrive or harddisk to create a Mona Lisa in a new, previously unseen way.
Just take for example Stable Diffusion and tell it to recreate the Mona Lisa. Maybe re-genrate a few times. You'll see it doesn't have the exact pixel values of the original image and you won't be able to get a 1:1 copy. If you look at a few outputs, you'll see it draws it from memory, with some variation. It also reproduces the painting being photographed from slightly different angles and with and without the golden frame around it. Once you tell it to draw it frowning or in anime style, you'll see that the neural network has learned the names of facial expressions and painting styles, and which one is present in the Mona Lisa. So much that it can even swap them without effort.
And even if neural networks can remember things very precisely... What about people with eidetic memory? What about the painters in the 19th century who painted very photorealistic landscape images or small towns. Do we now say this isn't original because they portrayed an existing village? No, of course it's art and we're happy we get to know exactly how things looked back then.
Well, you mentioned that it could reproduce the imagine 1:1 which is just my entire point. It doesn't matter what your thumb drive can't do.
And I guess the main point is that every pixel is used and trained without changes, making it kinda a copyright issue as some images don't even allow to be used somewhere and edited.
I think what i was trying to say is something different: It can not do that. only theoretically under specific circumstances, more a maths homework assignment, but not really in practice.